Study Designs and Applied Statistics Methods in the Clinical Research

Study Designs and Applied Stat istics Methods in the Clinical

Research

振興醫院教學研究部

董道興 副研究員

2008.2.26

spital

Part I: Basic Study Designs

研究設計的重要性

• 好的研究設計較能得到可信的研究結果

• 好的研究設計較能提供正確的因果關係

spital

研究設計之分

類 • 觀察性研究

1. 個案研究與描述性研究 (Case-series or Descriptive)

2. 橫斷研究 (Cross-sectional studies)

3. 世代研究 (Cohort or Prospective studies)

4. 病例對照研究 (Case-control or Retrospective studies)

5. 歷史性世代研究 (Historical cohort studies)

• 實驗性研究 (Experimental studies)

1. 控制性試驗 (Controlled trials)

A. Parallel or concurrent controls a. Randomized

b. Not randomized B. Sequential controls a. Self-controlled b. Crossover

C. External controls (including historical)

2. 無控制組

• 統合性研究 (Meta-analysis)

spital

觀察性研究

• 個案研究

1. 針對某些特定病徵的病人進行描述 2. 無對照組 (control group)

3. 無研究假設 (research hypotheses)

spital

• 橫斷研究 (Cross-sectional studies)

有病

沒病 研究樣本

Time

研究時間

No direction of inquiry Question: What is happening?

橫斷研究之特點

• 目的

1. 描述健康狀態 2. 分析因子與事件之相關

• 研究設計

同時間收集因子與疾病等變項

• 資料收集方法

1. 病歷紀錄 2. 觀察 3. 面訪及電訪 4.

問卷填寫

• 優點

1. 簡單易行 2. 資料收集範圍廣闊 3. 較具經濟性

4. 可短時間內得到結果 5. 較易獲得大量樣本

• 缺點

1. 無法判定因果關係 (causal relationship)

2. 推論易流於表面

spital

WHO (1999) criteria

Type 2 diabetics (n=1123)

Migrated out or death (n=152)

1994-1998

DR screening for Type 2 dia betics in 1999-2002

(n=971)

725 Type 2 diabetics received DR screening

Loss to follow-up (n=246) Screening

for type 2 diabetes (1991-1993)

Screening for diabetic retinopathy (1999-2002)

Economic evaluation of diabetic retinopathy screening (2003)

Utility survey

Willingness-to-pay survey

Sight year measurement

Economic evaluation

1. Cost-effectiveness analysis 2. Cost-utility analysis

3 .Cost-benefit analysis

Epidemiologic information 1.Prevalence of DR 2.Incidence of DR 3.Risk factors of DR 4.Natural history of DR

The Study Procedure

spital

有病

沒病

有病

沒病 暴露組

(exposed)

非暴露組 (unexposed) 研究世代

研究起始點

Direction of inquiry

Question: What will happen?

世代研究之特點

• 目的

探討暴露組和非暴露組之間發生率 (incidence) 是否相同

• 研究設計

1. 觀察世代可選擇一組或兩組 2. 長期觀察研究世代之得病情形

• 優點 1. 可計算發生率 2. 可了解因果關係

3. 可以調查稀有暴露世代的得病情形 4. 可減少選樣性偏差的發生

5. 可探討多重疾病 (multiple diseases)

• 缺點

1. 觀察時間過長 2. 費用較高

3. 失去追蹤導致結果產生偏差 (bias)

spital

• 主要測量指標

相對危險性

相差危險性

有病 沒病

有暴露 沒暴露

a b a+b

c d c+d

a+c b+d a+b+c+d

spital

• 成果評值 (Outcome assessment)

1. 功能狀態 (Functional status) 2. 生活品質 (Quality of life)

3. 滿意度 (Patient satisfaction) 4. 存活時間 (survival time)

5. 經濟評估 (Economic evaluation)

A.Cost-effectiveness analysis –CEA

C.Cost-benefit analysis –CBA

spital

spital

• 病例對照研究 (Case-control studies)

病例組 (cases)

對照組 (controls)

暴露 (exposed)

未 暴 露 (unexpose d)

Time 研究起始時間

Direction of inquiry Question: What happened?

暴露 (exposed)

未暴露 (unexposed)

病例對照研究法之特徵

• 目的

探討病例組與對照組其暴露分布是否相同

• 研究設計

選擇一組有病者為病例組 , 無病者為對照組 , 比較病例組與 對照組間過去的暴露經驗

• 病例組與對照組之選擇

1. 定義清楚所選擇的病例組個案

2. 選擇具有代表性的對照組個案

spital

• 優點

1. 迅速便宜

2. 所須樣本數較小

3. 適合於稀有疾病之研究 4. 可快速得到結果

5. 可以探討多重暴露對單一疾病的相關

• 缺點

1. 不易得到過去完整的暴露經驗 2. 不易選取合適的對照組

3. 僅能用估計的方式得到發生率 4. 時序性不易確定

5. 可能產生選樣性偏差 (selection bias) 或 回憶性偏差 (recall

bias)

病例組 a b a+b 對照組 c d c+d

暴露 未暴露

a+c b+d

1. 藉著比較 case 及 control 組的 exposed 與 unexposed 的 比值

(ratio), 即比較 a/b 和 c/d 就可知 exposure 和 disease 是否 有關

2. 若 case 組暴露率較 control 組高 (a/b>c/d), 表示暴露和 發病

有正相關

3.case 組的 exposure odds=a/b control 組的 exposure odds=c/d

則

如果 OR 值越高 , 表示暴露與發病越有正相關 .

spital

• 歷史性世代研究 (Historical cohort studies)

研究世代

暴露組 (exposed)

對照組 (unexposed)

有病

有病 沒病

沒病

Time 研究起始時間

Direction of inquiry

觀察性研究設計之比較

種類 過去 現在 未來 橫斷研究 Subjects selected,

data gathered on exposure status and outcome

病例對照研究 Information Case chosen

obtained on with outcome past exposure Controls chosen without outcome

世代研究 Subjects selected outcome and classified as measured

to exposure

歷史性世代研究 Subjects identified Subjects who were exposed

classified as to in the past exposure from

existing records Outcome measured

實驗性研究

(Experimental studies)

• Experimental studies that involve humans are called clinical trials

• Controlled trials are studies in which the experimental drug or procedure is compared with another drug or procedure, sometimes a placebo and sometimes the previously accept treatment

• Uncontrolled trials are studies in which the investigators’

experience with the experimental drug or procedure is described, but the treatment is not compared with

another treatment

spital

研究樣本

Time 有結果

有結果

無結果

無結果

對照組

介入 (intervention) 研究起始時間

• 實驗組與對照組

1. 接受實驗步驟的對象 所組成的團體稱為實驗組 , 未接受實 驗步驟者則稱為對照組

2. 實驗組的介入因素包括 : 治療性、預防性、干預性、社區 性試驗對照組則使用安慰劑 (placebo)

spital

1. 研究對象簽了同意書之後 , 根據 random allocation 來分組 , 通常採用隨機號碼表

2.Randomization=Random Allocation

指每一個研究對象被分到任一組的機率是一樣的 3. 以隨機分派的方法來分配實驗組與對照組可以提高兩組 的可比較性 (comparability), 而避免自我選擇 (self- selection) 的誤差 , 以建立資料分析結果的效度 (validity) 4. 隨機分派方式 :

(1) 簡單隨機分派 (simple randomization)

(2) 分區隨機分派 (Block randomization)

(3) 分層隨機分派 (Stratified randomization)

• 盲目程序 ( Blinding procedure)

1. 在評估實驗結果時 , 必須考慮研究者與參與者的 bias, 為了避免 bias, 最好的方法就是 double blind

2. 單盲程序 (single blind): 實驗者知道 , 被實驗者不知道被分到哪一組 雙盲程序 (double blind): 實驗者和被實驗者都不知道會分到哪一組 三盲程序 (triple blind): 實驗者、被實驗者和資料分析者都不知道

3. 雙盲或三盲程序有時不易做到 : (1) 副作用 (side effect)

(2) 有些 trial 是 home VS hospital

(3) 醫師必須知道 patient 是屬於哪一組 , 便於照顧病人的情況

• Randomized clinical trials

Provided the strongest evidence for concluding causation

• Nonrandomized trials

Do nothing to prevent bias in patient assignment

• Trials with self-controls

1.Only one group in which patients are assessed before and after the intervention

2.Hawthorne effect

• 交叉試驗 (Crossover study)

subjects meeting entry criteria

experimental subjects

controls

with outcome

without outcome with outcome

without outcome

experimental subjects

with outcome with outcome

controls

without outcome

without outcome

onset of study

intervention washout period intervention

Results from previous study

with outcome

with outcome

without outcome

without outcome

Time onset of study

intervention in subjects only

spital

y)

spital

統合性研究 (Meta-analysis)

• Meta-analysis uses published information from other studies and combines the results so as to permit an overall conclusion

• Meta-analysis is similar to review articles, but additionally includes a quantitative assessment and summary of the findings

• Meta-analysis is especially appropriate when the studies that have been reported have small numbers of subjects or come to different conclusions

spital

實驗設計評估

• How to evaluating the study design the medical literatures?

(1) How are subjects recruited?

(2) Are subjects randomly assigned to the study group?

(3) Is there a control group?

(4) Are appropriate therapies included?

(5) Is the study blind? Double blind?

(6) How is compliance evaluated?

(7) If some cases are censored, is a survival method

such as Kaplan-Meier or the Cox model used?

2. In a cohort study

(1) How are subjects recruited?

(2) Are subjects randomly selected from an eligible pool?

(3) How rigorously are subjects followed? How many dropouts does the study have and who are they?

(4) If some cases are censored, is a survival method such as Kaplan-Meier or the Cox model used?

3. In a case-control study

(1) Are subjects randomly selected from an eligible pool?

(2) Is the control group a good one?

(3) Are records reviewed independently by more than one person?

spital

4. In a cross-sectional study

(1) Are the questions unbiased?

(2) Are subjects randomly selected from an eligible?

(3) What is the response rate?

5. In a meta-analysis

(1) How is the literature search conducted?

(2) Are the criteria for inclusion and exclusion of studies clearly stated?

(3) Is an effort made to reduce publication bias?

(4) Is there information on how many studies are need to

change the conclusion?

Part II: Basic Statistics Methods

spital

Types of clinical data analysis

• Time-unrelated analysis

• Time-related analysis

Time-unrelated analysis

• Categorical scale

1. Proportion Z test & 

test 2. Logistic regression

• Interval scale

1. Independent T-test (Two-sample) 2. ANOVA

3. Linear Regression

spital

Time-related analysis

• Balanced repeated measurement

(1) Generalized Estimation Equation (GEE) (2) Mixed model

2.Interval scale (1) Pair T-test (2) Mixed model

• Survival analysis

Survival analysis

• Non-parametric method

Kaplan-Meier (Life-Table) method

• Semi-parametric method  Cox regression model

• Parametric method

spital

Two-Sample Problems

• Continuous variable 1.Independent T-test 2.Paired T- test

• Binary variable

1.Proportion Z Test or Chi-square test

2.McNemar’s Test

Independent T-test & Paired T-test

Example 1 : Evaluation of effectiveness of

community rehabilitation care (n=30) Pre Post

83 77 95 80

75 70 62 59

spital

Statistical evaluation

Mean

Pre Post 78.20 70.26 (SD=14.31) (SD= 18.04)

Paired T T=-3.09 P=0.0079

Independent T T=-1.33 P=0.19

Correlation & Repeated measurement

Paired T test enhance statistical efficiency by considering ... the correlation between Pre- and Post- (the same individual)

spital



test & Proportion Z test

Example 2 :

• The study on the assessment of mental health program

• Intervention: Experimental & Control

• Outcome : Favorable (F) & Unfavorable (U)

Two Data Sets

1. Data 1 :

F U

Test 40 20 Control 16 48

2. Data 2 :

F U Test 10 2 Control 2 4

spital

Statistical evaluation

• Data 1



  

Proportion test: Diff=P1-P2

=0.667-0.25=0.417 (95%CI: 0.26-0.58)

• The intervention group has a more significant favorable response than the control group

Fisher Exact Test

• Data 2:



=4.50 P=0.034 (right??)

Fisher’s Exact Test :

One-Tail P=0.057 Two-Tail P=0.107

spital

Correlated Data (McNemar’s Test)

Example 3 :

A match study on the efficacy of attending CRC screening with respect to health

promotion (matched by sex and age)

Outcome: whether to attend CRC screening

General 

test

+ -

+ 43 33 Attend

- 16 26



=3.70 P=0.055 (right??)

Health Promotion

spital

McNemar’s test

Health Promotion Yes

+ -

+ 27 16 43 No

- 6 10 16

33 26 59

₂₍₁₎

=4.54 P=0.033

Measuring agreement between two people

Example 4:

Two reviewers on mammography

Second

Yes No Yes 20 5 First

No 3 200

spital

Kappa statistics

Kappa =(Observed-Expected agreement) /(1-expected agreement)

=(O-E)/(1-E)

Kappa=0.70 (95% CI: 0.50-0.89)

How to interpret kappa?

Byrt (1996)

0.93-1.00  Excellent agreement 0.81-0.92  Very good agreement 0.61-0.80  Good agreement

0.41-0.60  Fair agreement 0.21-0.40  Slight agreement 0.01-0.20  Poor agreement

0.00  No agreement

spital

Mutivariate analysis

• Continuous variable

1. Linear regression

2. Mixed model (repeated measurement)

• Binary variable

1. Logistic regression

2. GEE model (repeated measurement)

Linear regression

Example 5 :

Is the effect of diet and exercise program on the SBP effective?

Independent variable: diet, exercise Outcome: SBP

spital

Multiple Linear Regression

Y= a+b1*diet+b2*exer+b3*diet*exer +

(Interaction term)

b’s SE T P-value

diet: 1.03 2.22 0.47 0.64

exer: 0.78 0.13 6.21 0.0001

diet*exer : 0.30 0.20 1.54 0.13

Logistic regression

Example 6:

A study was performed on 53 patients with prostate cancer to collect data on

several variables considered predictive of nodal involvement

spital

Definitions of predictors

• Age

• Acid: level of serum acid phosphate in King-Armstrong units (0.46-1.26)

• X-ray examination: (0=negative 1=positive)

• Grade: Pathological grade

(0=less serious 1=more serious)

• Size: size of tumor (0=small, 1=large)

• Nodalinv: Laprectomy results

(0=No involvement;1=Involvement)

Model selection

X-ray, Size and Acid are

three significant factors for predicting nodal involvement

spital

Model Prediction

    ^x

exp

x P exp

 β

 1

β

Odds Ratio (OR)=exp()

Probability prediction of nodal involvement

(risk classification)

• High-risk

Acid= 1.26 , X-ray with positive finding, and large tumor = 0.96

• Low-risk

Acid=0.48 , X-ray with negative finding, and small tumor=0.05

spital

Generalized Estimation Equation (GEE method)

• Indications: The regression model for correlated binary outcome

• Example 7

A clinical trial comparing two treatments

for a respiratory disease

Definitions of variables

• Outcomes: Respiratory status

(0=poor 1=good)four visits

• The main variable: Treatment

(1=Active 0=placebo)

• Other explanatory variables:

center, age, sex, and baseline respiratory status

spital

Characteristics of GEE model

Working correlation matrix

to accommodate correlation

between outcomes

Survival Analysis

• Outcome variable: survival time

• Starting pointsendpoints : Survival time

• Censoring problem :

i.e Right censoring: study ends or lost to follow-up

spital

Kaplan-Meier method

Example 8:

Prognosis for women with breast cancer:

A histochemical marker called Helix pomat

ia agglutinin (HPA) is used to assess whet

her tumor have already been metastasized

Definition of Variables

32 Patients

• Survival time: from treatment to death

• Censoring: 1=dead 0=censoring

• Prognostic factor: HPA (1=positive 0=negative)

spital

Cox regression model

• Proportional hazard model:

One of common multiple regression models for survival analysis.

• No baseline hazard estimation

• h(t,x

,x

,…,x

)=h

(t)exp(b

x

+b

x

+…+b

x

)

Importance of the Cox Model

• Provided the only valid method of predicting a time-dependent outcome

• Producing survival curves that are adjusted for confounding factors.

• The Cox model can be extended to the case of multiple events for a subject

• Cox model could estimate RRexp()

spital

Parametric regression models Accelerated failure time

(AFT) model

• To accommodates left censoring and interval censoring

• Distribution

1.exponential distribution 2.weibull distribution

3.log-normal distribution 4.log-logistic distribution 5.Gamma distribution

Other methods for multiple variables

• Discriminant analysis

It assumes that the independent variables follow a multivari ate normal distribution, so it must be used with caution if so me X variables are nominal

• Log-Linear analysis

All the variables, both independent and dependent, are mea

sured on a nominal scale

• Factor analysis

All variables are considered to be independent variables

spital

• Cluster analysis

The object is to determine a classification or taxonomic scheme that accounts for variance among the subjects

• Multivariate analysis of variance (MANOVA)

Involved multiple dependent variables as well as multiple independent variables

• Canonical correlation analysis

W hen both the independent variables and the outcomes

are numerical and the research question focuses on the

relationship between the set of independent & dependent

variables

Summary of conceptual framework for questions involving two variables

Independent variable

Dependent

variable Method

nominal nominal chi-square binary numerical t-test

nominal numerical one-way ANOVA

(more than two values)

numerical numerical regression; correlation

spital

Summary of conceptual framework for questions involving two or more independent variables

Independent variable

Dependent variable

Method

nominal nominal Log-linear

nominal and numerical nominal Discriminant analysis (two or more values)

nominal numerical ANOVA

numerical numerical Multiple regression

nominal with confounding numerical ANCOVA

nominal Mantel-Haenszel

Thank you for your attention!

spital